Structural EM for BNTWritten by Wei Hu (wei.hu@intel.com) 2002The attached files are SEM  codes and test applications. It is based on the junction tree inference engine, and use BK approximationto compute nodes' expected sufficient statistics.Test application "sem_learn1.m" learn a "sprinkler" network, which has 4discretenodes. "sem_learn1" can get exact answer like the original network in shorttime."sem_learn2.m" learn the "discrete1" network in BNT. It will run more than 1hour."sem_learn3.m" learn the "alarm" network. Actually I have not run it for awhole procedure,because it will take a so long time. This suggest that even with approximateESS,the structure learning for missing data, and for a moderate networks likealarmwill need parallelized.You can setup the SEM as follow,1.   copy the file "learn_struct_EM.m" into ../BNT/learning2.   copy file "mk_nbrs_of_dag_topo.m" into ../BNT/graph , the file iswritten by Diao Qian, based on       "mk_nbrs_of_dag.m"3.   copy "multiply_one_marginal.c" into ../BNT/misc, then use "mex" tocompile it4.   copy "sem_learn1.m" and  "sem_learn2.m" into ../BNT/examples/static       And optional, if you want to test the "sem_learn3.m", you can5.   copy "sem_learn3.m" and "mk_alarm_bnet.m" into ../BNT/examples/staticWhen you want to run SEM, better to run "installC" (../BNT) first,algorithms in C will be 4 or 5 times fasterthan pure MATLAB codes.Wei